Research map / Label-efficient and robust learning / Self-supervised and contrastive
Semantics-Consistent Feature Search for Self-Supervised Visual Representation Learning
Kaiyou Song, Shan Zhang, Zimeng Luo, Tong Wang, Jin Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantics-consistent feature search (local data to local feature contrast) (ablation) | DINO ResNet50 · ImageNet100 linear probing | linear probing top-1: 87.0 -> 89.2 | Improved |
| add multi-layer contrast to DINO instead of feature search (ablation) | DINO · ImageNet100 linear probing | linear probing top-1: 87.0 -> 87.4 | Improved |
| use soft feature search instead of hard ROIAlign crop (ablation) | SCFS · ImageNet100 linear probing | linear probing top-1: 88.1 -> 89.2 | Improved |
| add feature search to setting without multi-crop (ablation) | DINO without multi-crop · ImageNet100 linear probing | linear probing top-1: 83.7 -> 86.3 | Improved |
| apply feature search on Res2+Res3 instead of Res2 only (ablation) | SCFS · ImageNet100 linear probing | linear probing top-1: 86.1 -> 87.5 | Improved |
| apply feature search on Res2+Res3+Res4 instead of Res2+Res3 (ablation) | SCFS · ImageNet100 linear probing | linear probing top-1: 87.5 -> 89.2 | Improved |
| add SCFS to DINO on ResNet101 | DINO R101 · ImageNet100 linear probing | linear probing top-1: 86.3 -> 88.3 | Improved |
| add SCFS to DINO on ViT-S | DINO ViT-S · ImageNet100 linear probing | linear probing top-1: 80.4 -> 81.0 | Improved |
| add SCFS to DINO when pre-training on COCO | DINO ResNet50 · COCO detection after COCO pre-training | AP^b: 39.0 -> 40.9 | Improved |
| add SCFS to DINO (256 bs, 200 epochs) | DINO ResNet50 · ImageNet linear probing | top-1 accuracy: 73.0 -> 73.9 | Improved |
| add SCFS to DINO for COCO 2x schedule detection | DINO ResNet50 · COCO Mask R-CNN 2x | AP^b: 41.9 -> 42.1 (HSA 42.2) | Mixed |
| add SCFS feature search | DINO ResNet50 training cost · training time bs 256, 200 epochs | training time: 147h -> 192h | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OBoW — OBoW: Online Bag-of-Visual-Words Generation for Self-Supervised Learning (CVPR 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Beyond Prototypes: Semantic Anchor Regularization for Better Representation Learning (AAAI 2024 · similarity 0.85)
- Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.85)
- Self-Supervised Representation Learning with Meta Comprehensive Regularization (AAAI 2024 · similarity 0.84)
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 2025 · similarity 0.84)
- Self-Organizing Visual Prototypes for Non-Parametric Representation Learning (ICML 2025 · similarity 0.84)
- Progressive Feature Self-Reinforcement for Weakly Supervised Semantic Segmentation (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
